Found while verifying v1.8.7 from inside a fresh container: the baked mempalace
snapshot is read by no host on this fleet. devbox-skill-reconcile repoints
~/.agents/skills/mempalace at the mounted live clone (the v1.8.5 fix working as
designed), and all four compose stacks mount a workspace containing the
skillset. So the phrase canary that blocked v1.8.7's first tag polices a file
nobody opens, while the drift that could actually mislead an agent — a git pull
nobody ran in /workspace/skillset — was invisible from inside the container and
is invisible to CI by construction.
Record provenance instead of policing it, and move the check to where the
skillset actually is:
- Dockerfile.variant: ARG SKILLSET_SNAPSHOT_REF (the claim) + a sha256 of the
shipped bytes measured in the manifest layer (the fact), as manifest siblings
rather than components{} members, plus an OCI label. An ARG default, not a
CI-resolved output: no credential for the private skillset, no change at any
of the four variant build call sites, and a local docker build records what CI
does. Variant-only, so no base rebuild — check-base-hash.sh scans
Dockerfile.base alone, verified by running it.
- pi-devbox-version: a skills: section naming baked vs live <repo> @ <sha> per
vendored skill, and for mempalace whether the live copy is identical to the
baked fingerprint, at the same commit with uncommitted edits, or divergent.
entrypoint-user.sh passes the new --no-skills, because the banner prints
before the links exist and long before the reconcile runs.
- scripts/vendor-mempalace-skill.sh: refresh the file and rewrite the ref
together (a cp without an ARG bump makes the manifest lie, which is worse than
anonymity); --check verifies the claim against a real clone.
- 5 new smoke assertions (78 -> 83), mutation-tested through the real sh -c
path: 6 fabricated manifests, where a well-formed hash of the wrong file
proves the two manifest assertions are not redundant; the all-baked reporting
test verified to FAIL against a live-skillset environment.
Reviewed mid-flight by pi@emb-7kj4vr4g over the logstream (correlation
skillset-vendor-drift), which retracted its own earlier recommendation of a
build-time byte-compare against skillset HEAD and supplied the better framing:
the invariant is NON-CONTRADICTION, not currency. Byte parity on a fallback
would have cost a resync commit plus a ~67-min base rebuild for each of the four
skillset commits pushed in one evening. Its warning also found a real bug here:
the script now CONSTRUCTS the snapshot from `git show HEAD:<path>` instead of
copying the working tree, because a clean `git diff` says nothing about an
untracked file — the one input the first draft would have recorded a false ref
for. Tested: untracked, unstaged and staged-but-uncommitted all refuse, atomically.
Also fixes three stale in-repo markers of the same class the canary belongs to
(true when written, silently false at release): two dangling "Unreleased"
pointers and a typst line still marked Unreleased five releases after v1.4.0.
joakimp
e070e0bcbf